Sampling strategies for forest aerial detection survey in Colorado
Date
2015
Authors
Ha, Anh Quang, author
Reich, Robin M., advisor
Jacobi, William R., committee member
Lundquist, John E., committee member
Khosla, Rajiv, committee member
Journal Title
Journal ISSN
Volume Title
Abstract
Aerial detection survey (ADS) has been commonly employed in forest surveys in the United States for detecting forest damage and monitoring forest health. In Colorado, ADS by USDA Forest Service has conducted annual 100% census of government forested land for more than 20 years with the goal of achieving information about forest damage due to different causal agents and disorders. Sketchmapping has been commonly employed in ADS with the goal of detecting and documenting on maps mortality, defoliation and other visible forest change from aircraft. At medium and large scale, sketchmapping is a suitable technique for forest monitoring that provides valuable information in forest health. This dissertation deals with data of forest area damaged by five causal agents mountain pine beetle, spruce beetle, western spruce budworm, pin engraver, and Douglas fir beetle and two disorders subalpine fir mortality and sudden aspen decline. The combined areas damaged by all causes were also considered. Data were downloaded from ADS in Colorado from 1994 to 2013 as polygon shapefiles with associated information such as causal agents or disorders, area damaged, and type of forest. The goal of my dissertation was to identify an appropriate sampling strategies to archive good estimates of total area damaged, to decrease survey cost, and to increase safety by reducing the amount of flights. To approach this goal, four sample designs for estimating total area damaged caused by various causal agent were evaluated: simple random sampling, stratified random sampling, probability proportional to size, and non-alignment systematic sampling. A GIS layer of 150 transects covering Colorado’s forestlands was developed and represented the sample unit for my study. Each transect was 3.2 km wide and 625 km long and was numbered from 1 to 150 from south to north. Each sample design was evaluated using eight sample sizes (10, 15, 20, 25,30, 35, 50, and 70) and applied to the seven damages and the combined damaged area. The statistical properties were evaluated to determine the optimal sample design for estimating area damaged caused by different causal agents. The spatio-temporal characteristics of area damaged that influence precision and accuracy of estimate were considered. Most of the damaged forest areas by single causal agents and disorders showed aggregated spatial patterns; whereas the combined damaged areas were uniformly distributed across the landscape. A loss plus cost function was employed to determine the optimal sample size for each sample design and analyzed for the cost advantage of alternative sample designs. We found that stratified random sampling was the most optimal sample design by producing the highest percentage of unbiased estimates of total area damaged and the smallest variances. The next best sampling designs were simple random sampling and probability proportional to size. The non-alignment systematic sampling was the worst for estimating total area damaged both for individual causal agents and disorders and all causal agents combined. The optimal sample size varied by sample design and causal agents and disorders as well as the level of confidence. Optimal sample size increased with increasing variability in the population and as the desired level of confidence increased. Larger samples were required to simultaneously provide estimates for multiple causal agents and disorder with reasonable levels of precision when compared to a single causal agent. Stratified random sampling was the most cost effective when compared with other sample designs. For example, the cost advantage of stratified sampling over random sampling for estimating the damage from subalpine-fir mortality was $85,000 per year. In contrast, PPS sampling had a cost disadvantage of -$13,000 per year when compared with simple random sampling and -$95,000 per year when compared with stratified sampling for estimating the total damage from all causal agents combined at the 0.95 level of confidence.